Cost-time trade-off efficient workflow scheduling in cloud

Highlights

• The proposed scheduling algorithm aims at minimizing user deadline and budget.

• A range of instance types that best suit the workflow execution is determined.

• A fine-grained cost-time trade-off factors produce the most viable schedule.

• The types and the number of VMs, and the trade-off factors are very important.

Abstract

Cloud computing has become a promising solution for scientific workflow applications, due to its various assets. Workflow scheduling is a well-known NP-complete problem, so difficult to solve since there are no optimal solutions. Several workflow scheduling works aim at optimizing the makespan and the budget. However, more investigations are needed for appropriate resources chosing in the large set of instance types offered in cloud environments. This paper proposes a new scheduling algorithm called Cost-Time Trade-off efficient Workflow Scheduling (CTTWS), which consists of four main steps: task selection, Implicit Requested Instance Types Range (IRITR) evaluation, spare budget evaluation and VM selection. The IRITR evaluation is a novel scheduling concept, which aims at determining a range of VMs instance types that best suits the workflow execution, in order to avoid overbidding and underbidding that may lead to budget and deadline violation respectively. Compared to previous work, the results of simulations prove the effectiveness of our approach, especially when there is a large variety of instance types. This confirm that paying attention to the type of resources and their number is vital.

Introduction

Almost all scientific areas are nowadays more complex and relies on the analysis of large scale data sets, it is therefore required to use an automated management process in a scalable way. Scientific workflows have emerged as a suitable way of describing and structuring parallel computations and the analysis of large scale data sets [1]. Their successful use enhanced scientific advancements in various fields such as biology, physics, medicine, and astronomy [1], [2]. Complex experiments based mainly on the analysis of large-scale data sets, which sometimes require high computing power [1], are gradually exploiting the assets of commercial clouds [3], [4], [5], [6], [7].

Cloud providers offer several types of services, the main ones been the Software as a Service (SaaS), the Platform as a Service (PaaS), and the Infrastructure as a Service (IaaS) which is the principal service that benefits Workflow Management Systems (WMSs). The SaaS is offering dedicated software to cloud end users that are accessible on the Internet via a browser; The PaaS is providing the platform and the necessary Information technology (IT) environment for developers to implement their various services and applications on the Internet; while the IaaS is delivering virtualized resources called Virtual Machines (VMs) for lease to support users operations.

We can enumerate among the advantages of the Cloud: (i) its flexibility and elasticity; (2i) its pay-per-use billing model, which is helpful to avoid upfront expenses for personal or dedicated resources purchasing [5], [7]; (3i) its reliability and fault tolerance; (4i) and its variety of resources type (instance type) going from general purpose to specialized resources such as GPUs [8].

Workflow Scheduling is a well-known NP-complete problem [9]. The level of difficulty of workflow scheduling is also related to the number of factors to consider, as the user defined Quality of Service (QoS) like deadline, service cost, and so forth. Several researches on workflow scheduling for Heterogeneous Computing Systems (HCS) aim to minimize deadline and/or budget [10], [11], [12], [13], [14], [15], [16].

However, in regard to the above-mentioned complexity of both cloud environment and workflow structure, it is essential to design scheduling algorithms tailored for scientific workflows in order to take more advantage of clouds assets. Although the cloud has several advantages, like for instance its flexibility and elasticity, inefficient usage of resources and high computing costs may result if inadequate scheduling and provisioning decisions are made [17]. For example, it is very important to determine the types and the number of appropriate resources [18] and insure a good workload management [19] in order to avoid energy wastage and Service Level Agreement (SLA) violation when running workflow tasks.

In an unknown and diversified market, the assistance of a good sales consultant is important to tailor the quality of purchases to the customer needs and budget. In a cloud, this problem of optimizing the budget-quality ratio becomes the problem of optimizing the execution time and the computing cost according to the user-defined budget and deadline, and the role of a good sales consultant is played by a good scheduling algorithm.

In this paper, we propose a new workflow scheduling algorithm that aims to optimize execution time and processing cost, the Cost-Time Trade-off effective Workflow Scheduling (CTTWS). The CTTWS scheduling algorithm uses a novel concept, the Implicit Requested Instance Types Range (IRITR) evaluation, to determine a range of VMs instance types that best suits the workflow execution, in order to avoid overbidding as well as underbidding that may lead to budget and deadline violation respectively. Thereby, root tasks are executed on relatively fast instances that speed up execution, with the ability to be reused, and no task uses a slower instance than those in the IRITR. Our algorithm also uses new trade-off factors between time and cost to determine the most viable schedule, and uses this to get the most appropriate type of VM instance to provision. In this work, our trade-off function and its related issues, namely task selection and sparse budget evaluation, are based on a fine granularity approach, compared to their counterparts in the Budget Deadline Aware Scheduling(BDAS) algorithm [16], one of the most recent published work related to our goal and conditions, that rely on a big granularity approach.

We conducted experiments using WorkflowSim [20], an extension of CloudSim [21] for investigating workflows. Five well-known scientific workflows generated by the Pegasus workflow generator [22], namely Montage, Epigenomics, Cybershake, SIPHT and LIGO, each consisting in size of 50, 100, 200, 500 and 1000 tasks. We compare CTTWS and BDAS [16], stemming from their cost ratio, time ratio and success rate. This study shows that CTTWS is more effective than BDAS, and outperforms BDAS up-to 38.4% in terms of successful scheduling.

The remaining sections of the paper are organized as follows. Section 2 presents related work. Section 3 defines the workflow scheduling problem. Section 4 describes the CTTWS algorithm. Section 5 presents the experimental results for CTTWS and BDAS, and compares them. Section 6 concludes the paper.